The trillion-dollar bet that free cash flow can wait

TWO pieces of data released on Monday told a story of an economy that is not slowing, and a company that is spending as if it has nothing to lose. The ISM manufacturing index rose to 55.6 in July from 53.3 in June, the strongest expansion since May 2022, with factory employment returning to growth for the first time since January 2025. Amazon's stock, meanwhile, hit a record high of $286.20, pushing its market value through $3 trillion for the first time. The two events share a theme. The American economy is humming. The question is who gets to keep the receipts.

Amazon's second-quarter results, reported on Thursday, were a lesson in accounting dualities. Revenue reached $200.6 billion, above the $197 billion consensus. Earnings per share of $5.75 demolished the $1.82 estimate. The catch: roughly $53.4 billion of pre-tax gains, primarily on Amazon's investment in Anthropic inflated the bottom line. Strip those out and EPS sits at about $1.95, still above consensus but hardly the kind of number that justifies a $3 trillion valuation on its own. More disquieting was free cash flow, which turned negative for the trailing twelve months at minus $7.6 billion, after standing at $18.2 billion a year earlier. Capital expenditure hit $169 billion over the same period, up $66 billion, as Amazon races to build the data centres that artificial intelligence demands.

The $3 trillion club now has five members: Nvidia, Alphabet, Apple, Microsoft and, as of Monday, Amazon. To join that group while bleeding free cash flow is unusual. It signals that investors have placed a bet, not on what Amazon earns today, but on what it will earn once its infrastructure is built. CEO Andy Jassy framed the cash-flow squeeze as a timing problem. Data centres require capital roughly two years before servers can be installed. The servers themselves break even in under three years, then run for five or six. Most of Amazon's AI capacity, he said, is contracted for terms of five years or longer. If that contract pipeline holds, the spending is not gratuitous. It is a bridge.

The bridge has a destination. Amazon Web Services grew 37 per cent in the second quarter, its fastest pace since 2021, to $42.2 billion in revenue, or a $169 billion annualized run rate. Operating margin in the cloud unit rose to 39.4 per cent from 32.9 per cent a year earlier, evidence that AI spending is starting to convert into profit, not just scale. AWS's own AI and custom-chip businesses each passed $25 billion in annualised revenue, more than doubling from last year. Amazon is no longer merely hosting the AI revolution. It is selling the tools for it: model-building services, inference infrastructure, agent-building platforms and, increasingly, the chips themselves. The deal with Meta for hundreds of thousands of Graviton chips is a sign that the silicon play is moving beyond internal use.

To be sure, the market's sanguinity about negative free cash flow has a precedent. Amazon endured years of thin margins and heavy capex during the original AWS buildout, and patient shareholders were rewarded. The pattern is seductive: spend voraciously, build an indispensable platform, watch margins expand as the fixed cost base spreads over ever-larger revenue. But AWS was built in an era of cheap money. The current cycle plays out with borrowing costs substantially higher and the Federal Reserve constrained by an economy that shows no sign of weakening. The ISM employment index rose to 52.8 in July, its highest reading since August 2022. That is good for workers. It is less comforting for a company carrying, by some estimates, over $119 billion in long-term debt and raising its 2026 capex guidance from $200 billion to $220 billion.

The deeper structural question is whether AI infrastructure will generate the same returns as the original cloud buildout. AWS succeeded because it solved a real problem at a time when enterprises were desperate to move off legacy systems. The addressable market was large, the competition was weak and the pricing power was enormous. AI infrastructure is less certain. The customers are fewer-the frontier model builders and the enterprises that want to run them. The spending is concentrated. And the risk of overbuilding is real: if demand does not grow as fast as the hyperscalers predict, the data centres that Jassy envisions monetising for thirty years become stranded assets.

The competition is also intensifying. Microsoft's Azure and other cloud services are growing at 43 per cent and have passed $100 billion in trailing revenue. Google Cloud surged 82 per cent to almost $25 billion in the quarter, with nearly $78 billion on a twelve-month basis. Amazon remains the largest cloud provider by a comfortable margin, but the gap is narrowing at the margins that matter-AI services, where Google's growth rate is the highest and Microsoft's integration with OpenAI and its own models is deepening. Every dollar Amazon spends on data centres is a dollar that does not accrue to a dividend, a buyback or a buffer against the day when AI growth disappoints.

Yet the case for Amazon is not obviously weak. The company's advantages in cloud are structural: the broadest capabilities, the strongest security posture, and the incentive alignment of running AI workloads alongside core applications and data. Its chip business-now over a $25 billion run rate, up from $20 billion three months ago-is a hedge against both Nvidia's pricing power and supply constraints. Advertising, another fast-growing segment at $76 billion over the past year, is highly profitable and benefits from the same ecosystem lock-in that protects AWS. The retail business, for all its lower margins, generates the scale and logistics network that no competitor can match.

The risk is not that Amazon will fail. It is that the $3 trillion valuation assumes everything goes right: AI capex monetises on schedule, interest rates fall rather than rise, competition does not erode margins and the Anthropic-style gains do not repeat. The market has decided that free cash flow can wait. That may be correct. Or it may be the kind of confidence that looks sensible until it isn't.

For the broader economy, the ISM data suggests the Federal Reserve has some room to be patient. Manufacturing is expanding, employment is growing, new orders are solid. The tariff risks that panelists flagged in previous months have not yet derailed the recovery. A strong economy is good for Amazon's retail and advertising businesses, and for the enterprises that buy cloud services. It is also bad, in the sense that it keeps financing costs high and limits how much the central bank can ease.

The aim should be for policymakers to keep rates where they are until inflation is clearly subdued, and for investors to treat Amazon's valuation as a bet on execution, not a verdict on inevitability. Better to price in the risk of capex disappointment than to assume it away.